Research Projects

Optimizing model architectures and pre-training approaches to improve automated hematological image analysis.

Deep Learning Strategies for Accurate Bone Marrow Cell Classification

The paper investigates how different deep learning models and pre-training strategies can improve automated classification of bone marrow cell images. It observes that models pre-trained on large, out-of-domain datasets like ImageNet perform markedly better, achieving substantial gains in precision and recall over models trained from scratch. Visualization with Grad-CAM shows that the networks often rely on meaningful morphological features, supporting their reliability in clinical contexts. Overall, the work establishes a strong performance benchmark and highlights the need for more high-quality data, especially for rare or difficult cell types.